Research On Chinese Continuous Speech Recognition In Noisy Environment

At present, Chinese continuous speech recognition has got a good performance in pure speech environment.But recognition rate is still low because of the influence of the noise in nature environment. It can not be used in actual occasion. So continuous speech recognition in noisy environment is always the hot spot and the difficult spot of the speech signal research.HMM is one of the most widely used models in continuous speech recognition. But the training of the model is often completed in quiet laboratory environment. Therefore, the trained model can not effectively describe the characteristics of the actual speech signal in natural environment. Because the inference of the noise, the accuracy of the voice activity detection is greatly reduced. So the system recognition rate is greatly reduced. Besides, the selection of the recognition unit and the context dependent of the model both have greatly affected the recognition rate in Chinese continuous speech recognition.In order to improve the performance of the speech recognition system in noisy environment, some speech recognition algorithms are researched according to the basic principles of continuous speech recognition and the characteristics of the noisy environment. This thesis has accomplished these tasks listed as follows:(1)A speech enhancement algorithm based on improved spectral subtraction is researched. This algorithm solved the "music noise" problem of the basic spectral subtraction. And an voice activity detection algorithm based on Adaptive sub-band spectral entropy is used. We combined it with improved spectral subtraction to improve the accuracy of voice activity detection in noisy environment further(2)Some speech feature extraction algorithms are introduced, such as:LPCC which reflects the information of sound channel and the MFCC with anti-noise performance. And a pitch detection algorithm based on LPC and Normalized Cross-Correlation is proposed. This algorithm improved the performance of pitch detection in noisy environmen (3)The basic principle of HMM model is discussed and the Forward-Backward algorithm, Viterbi algorithm and Baum Welch algorithm of the HMM are introduced. Meanwhile, the initialization of HMM, the overflow problem and the combination of HMM are explored in this thesis.(4)According to the training of acoustic models and pronunciation coordination problem in Chinese continuous speech, the selection of recognition unit, the embedded training of acoustic model and the context dependent tri-phone model are researched. At last, a series of experimental analysis is given based on HTK.